search

zhuoxiao-chen / ReDB-DA-3Ddet

[ICCV 2023] Revisiting Domain-Adaptive 3D Object Detection by Reliable, Diverse and Class-balanced Pseudo-Labeling
Apache License 2.0
46 stars 8 forks source link

How to choose pre-trained model? #4

Open csj777 opened 9 months ago

csj777 commented 9 months ago

Thanks for your outstanding work. In your supplementary about selecting pre-trained model, you say"For selecting the pre-trained models, we simply opt for the one with the lowest source risk." So what does the lowest source risk mean? Does this mean choosing the model with the best performance in the source domain?

zhuoxiao-chen commented 7 months ago

Hi, based on our observation, most source pretrained models will reach the best performance (lowest source risk) in the source domain, at very late epochs (checkpoints), meanwhile, achieving relatively ok performance on the target domain. So we normally choose the last checkpoint. However, for some tasks like Waymo to KITTI, the performance of final checkpoint could be very bad, if that is the case, we simply follow previous work (https://github.com/CVMI-Lab/ST3D/blob/master/docs/GETTING_STARTED.md#train-the-pre-trained) to select checkpoint which performs good on target.